HY-Image 3.5: What Tencent's Image Model Is, and What the Preview Label Actually Means

There are two cheap ways to waste a week on HY-Image 3.5, and both start with its name. The first is to treat it as a shipped product with a stable identifier you can pin a pipeline to. The second is to treat the word "preview" as a synonym for "not yet real" and file the model under things you will read about later. Neither reading survives contact with what the publisher actually documents. HY-Image 3.5 is Tencen​t's current image generation model, published by Tencen​t itself through its Hunyuan line and served on Tencen​t Cloud, and what is unresolved about it is not whether it works — it is what you are allowed to call the thing you are calling, and what identifier to write into a config file. That is the question this page answers, and it is the question a large share of the people searching for this model are already asking.

Who publishes it, and what it currently is

Tencen​t publishes it. The model sits inside the Hunyuan family, and Tencen​t's own product hub lists image generation as one of the line's core capabilities — 混元生图, described on Tencen​t's product page as built on the Hunyuan text-to-image model and offering strong Chinese comprehension alongside image generation. The commercial surface is Tencen​t's own cloud: the image generation model is a product on Tencen​t Cloud, reachable through the platform's TokenHub API surface, and documented there alongside the rest of the company's AI services.

The release state is the part worth being exact about, because it is the part the naming obscures. Tencen​t ships this line preview first and official version later, and the current generation is in the first of those two states. The identifier in circulation is hy-image-v3.5-preview. There is no second, stable, non-preview identifier for this generation sitting beside it, and Tencen​t has published no date for when one will appear.

Why this page leads with the preview question

Because the demand says so. Our own first-party search data — Googl​e Search Console, exact query match, the four weeks ending 2026-10-07 — puts this model at 430 impressions and 10 clicks across every spelling that reduces to its name. Broken out, that is 138 impressions for "hy image3.5" and 128 for "hy image 3.5" — and then 115 for "hy image3.5 preview", 45 for "hy image 3.5 preview", and 4 for the hyphenated identifier "hy-image-3.5-preview".

Read those numbers as a description of what people type, which is all a search console tells you. More than a third of the demand for this model's name arrives with the word "preview" attached to it. Those readers are not asking whether Tencen​t announced something. They are asking which designation they are dealing with, and the honest answer is that both the short name and the preview name point at one artefact: the preview generation Tencen​t is serving right now.

So, resolved: HY-Image 3.5 and "HY-Image 3.5 preview" are the same model at the same release state. The preview designation is current, not historical. The long name is the accurate one, and the short name is what the model gets called in coverage and in tool integrations because nobody types the qualifier when naming a thing they use.

What it accepts, what it returns, and the size envelope

Stated as Tencen​t states it, and flagged where the vendor's material is thinner than the coverage around it:

• Input — a text prompt, and for the editing direction, reference images. The attributable figure for reference images is up to five per request, which is the number carried consistently across Tencen​t's product material. At least one aggregator has published twenty; five is the figure we can attribute, and we are not upgrading it.

• Modes — text to image, image to image, and multi-turn editing, where an edit is applied against the accumulated context of the conversation rather than each request starting clean.

• Output — images up to 2K, at a set of common aspect ratios rather than a single square. Tencen​t's own words for the ceiling are "up to 2K".

• Billing — on delivered output images only. A request that fails or is filtered is not the thing you are charged for, which matters when you are sizing a workload rather than a demo.

One number in circulation does not survive the vendor's own material. At least one outlet reported a 4096×4096 output specification. Tencen​t says "up to 2K", and every other source agrees with Tencen​t. Treat the 4K figure as an error rather than as a roadmap for a later release — it is a correction, not an announcement.

Text rendering and editing behaviour

Two claims carry the model's behaviour, and both come from the publisher.

The first is lettering. Tencen​t claims improved accuracy on text rendered inside the image, in Chinese and in English both. That has been the standard weak point of the whole category, and it is the kind of claim that is easy to test on your own material and hard to settle from the outside, because nobody publishes a public text-rendering score for this model.

The second is consistency, which is the word Tencen​t leads with. The claim is that a subject survives repeated edits and re-renders without drifting — that editing refines the image rather than restarting it, which is the difference between a tool you use for one shot and a tool you use across a series of them. Paired with the five-reference-image envelope above, that is the documented basis for anything you build that needs the same character to reappear.

Both are vendor claims. Neither has been reproduced outside Tencen​t, and no independent board scores this model on text rendering at all.

The character-sheet question

Our own coverage of this model already talks about character sheets, so it is worth being precise about what the publisher documents and what it does not. What Tencen​t documents is multi-reference conditioning — you supply several reference images — combined with the consistency claim above. What Tencen​t does not document is a named "character sheet" feature, a sheet builder, or a per-character asset object. No such thing appears in the vendor's material for this model, and it appears nowhere in the documentation of the ComfyUI nodes that call it either.

So the capability is real but its name is ours, not theirs. If you want a character to hold across a run of images, the documented pieces you are assembling are: supply the reference images, up to five per request, and rely on the model's consistency behaviour across repeated edits. Whether that amounts to a production character pipeline is a judgement about your own shot list, not a specification Tencen​t has published.

Licence and access route, as the publisher documents them

There is no open-weights route here, and that is a documentation fact rather than a limitation being read into the material. Tencen​t publishes no downloadable checkpoint for this model, no licence for local use, and no offline path. Every documented way to reach it is either Tencen​t's own API or one of Tencen​t's own products.

The API is the route for anything at volume. The model is served on Tencen​t Cloud under the image generation product, reachable through the platform's TokenHub API surface, and it carries a published rate card: ¥0.15 for a 2K image on the domestic endpoint and US$0.024 for the same call internationally, both billed on delivered output images only. That international figure is corroborated independently — the model appears on the Artificial Analysis image boards at $24 per 1,000 images, which is the vendor's rate restated per thousand rather than a third party's own estimate.

Beyond the API, Tencen​t has put the model into its own first-party surfaces: Yuanbao, the consumer assistant; the ima knowledge tool; and the creative and productivity products Miora, WorkRally, OnSolo and WorkBuddy. It is also wired into Tencen​t Cloud's media stack through Media Processing Service and Video on Demand — which tells you the intended buyer includes teams already transcoding and editing video on Tencen​t infrastructure, not only developers building image features from nothing.

Third-party tooling exists but is thinner than it looks. ComfyUI shipped partner nodes for this model, and the important detail is what kind of node they are: they call Tencen​t's API. There are no local weights behind them, no checkpoint, no GPU path. "There is a ComfyUI node" and "I can run this on my own hardware" are different statements, and for this model only the first one is true.

Where the independent evidence actually puts it

This is the part the vendor's material cannot supply, and reading it carefully is the difference between a useful reference page and a press release. Artificial Analysis carries the model on two boards, under the name HunyuanImage 3.5 (Preview) — the board registers it under the preview designation too, which is its own corroboration of the naming above. Read today, 2026-10-13:

• Text to image — 68th on a board of 168 models, at 976.02 Elo, with a 95% confidence interval of 967.02 to 985.02, priced at $24 per 1,000 images. That places it in the middle of the field. The board's leader, GPT Image 2.5 Sunburst (max), sits at 1198.36 Elo; GPT Image 2.5 Flare (max) is at 1190.81, GPT Image 2 (high) at 1171.83, Nano Banana 2.1 at 1160.29, and Grok Imagine Image 2.0 at 1155.4. Tencen​t's previous generations on the same board: HunyuanImage 3.0 Instruct at 995.91 and 50th, HunyuanImage 3.0 (Tencen​t Cloud) at 983.72 and 64th, HunyuanImage 3.0 at 943.61, and HunyuanImage 2.1 down at 887.35.

• Image editing — 16th on a board of 81 models, at 1087.14 Elo, with a 95% interval of 1078.14 to 1096.14, again at $24 per 1,000 images. That places it in the top fifth of that board. The leaders are GPT Image 2.5 Sunburst (max) at 1184.26, GPT Image 2.5 Flare (max) at 1164.63, and MAI-Image-2.6 at 1140.28. Tencen​t's earlier generations: HunyuanImage 3.0 Instruct at 1068.07 and 23rd, HunyuanImage 3.0 (Tencen​t Cloud) at 1030.31 and 34th.

The gap between those two placements is the single most informative thing on this page, and it is not a contradiction. It says where the model is competitive: editing, where it lands near the top of a smaller, newer board, is where the consistency claim is being cashed. It says where it is not: on raw text-to-image preference it sits among the mid-field, well below the models at the top, and below where a reader arriving from a "+30% over the previous generation" headline might expect it to be. Both readings come from the same independent board, measured the same way, on the same day.

Two caveats belong with those numbers. Elo intervals are wide enough that adjacent ranks are not a meaningful ordering — a few Elo points between neighbouring rows is noise, not a finding. And matched-effort, matched-harness comparison between this model and the leaders above does not exist publicly; the boards rank preference votes, not a controlled bake-off.

How a reader actually uses it today

Start with the route that matches your intent. If you are evaluating, Tencen​t's first-party products are the cheapest way in, and the free two-week window that ran on Miora, WorkRally and OnSolo at launch closed on 7 October 2026 — so budget for the API from the first call rather than planning around a trial. If you are building, the API on Tencen​t Cloud is the only path with a rate card, and the meta-token surface is where the documentation lives.

Then write the generation call so the model name is a configuration value and not a constant. That is not stylistic advice for this model in particular; it is the discipline that keeps a preview-to-official transition — or a price revision, or a switch to a different generator entirely — from being a code change. A preview identifier is, by definition, the thing most likely to be superseded, and the cost of being ready for that is one line in a config file.

Where that discipline pays off is at the routing layer. OrcaRouter runs on a 0% markup model — the provider's list price passed through unchanged, so a vendor price revision lands the same day rather than waiting on a repricing cycle — and one API key covers more than 200 models, with automatic failover across providers and a routing DSL for pinning a specific model per request. That is what lets an image generator be a routing decision instead of an architectural one, whichever generator you point it at.

The short version, for a reader who arrived from a search box and wants the state of play: HY-Image 3.5 is Tencen​t's current image model, it is in preview under the identifier hy-image-v3.5-preview, it takes text and up to five reference images and returns images up to 2K, it is billed per delivered output image at a published international rate of $0.024 for a 2K image, it has no open-weights or offline route, and on the independent boards it is a top-fifth editing model and a mid-field text-to-image model at the same time. The preview designation is the open question, and the thing that would close it is a stable identifier with a date attached.